A claims adjuster types into an AI tool: “Summarize this claim report.” The AI returns three fluent paragraphs. They read well, but the date of the incident is incorrect. It wasn’t in the report at all—the AI added it. This is a scenario that could happen in this or a similar way. Whether it happens also depends on how the request is phrased. This article explains what information AI needs to provide a useful response—and what it must never be given.
A prompt is the instruction you use to have a language model perform a task. It’s best to think of it like handing off a task to a new colleague: attentively and quickly, but without any knowledge of your company, your conditions, or the specific process. Several common guides from providers outline the same basic principles: clearly state what you want, provide context such as the target audience and purpose, set boundaries, and specify the desired structure of the response. In the insurance industry, where wording and deadlines matter, these factors help determine whether a draft is usable or needs to be rewritten.
An example, fabricated and anonymized:
Before: “Summarize the damage report.”
After: “You assist with claims processing for homeowners’ insurance. Below is an anonymized claim report regarding a possible water pipe leak. Extract from it the date of the damage, location of the damage, sequence of events, affected items, and the stated amount of damage. Present the results in a table. Do not add anything that is not in the text. If a detail is missing, write ‘not specified.’ Do not make any statements regarding coverage.”
The second prompt is longer. However, it makes it clear what is expected, and the decision on coverage remains with the claims adjuster, where it belongs.
When copying text, names, addresses, or diagnoses can easily slip into the prompt. A guideline from the State of Berlin on working with AI chatbots clearly states: Personal or sensitive data—such as information about customers or trade secrets—does not belong in the query. Simply objecting to the use of data for training purposes is therefore not sufficient if personal data is entered. The document is aimed at the Berlin administration, but the basic principles are also instructive for businesses. According to Article 9 of the GDPR, health data falls under the categories of data requiring special protection, which often plays a role in cases of personal injury and in accident and health insurance. The in-house data protection officer should clarify which tools may be used with which types of data.
Language models can make claims that are incorrect. The BSI explicitly identifies a lack of factual accuracy—so-called “hallucinations”—as well as the lack of reproducibility of outputs as risks associated with generative AI. In everyday life, three habits can help: keep tasks narrow (summarize and organize rather than reinvent), cross-check numbers, data, and deadlines against the original source, and treat every result as a draft. The BSI also advises conducting a risk analysis for the specific use case before deployment.
Anyone who has found a template that works should share it. A shared collection of verified prompts for each task—such as evaluating damage reports, drafting customer letters, or comparing terms and conditions—prevents everyone on the team from having to start from scratch. This ties in with a requirement that many companies are not yet aware of: As of February 2, 2025, Article 4 of the AI Regulation requires companies that use AI to promote their employees’ AI literacy. Since the revision by Regulation (EU) 2026/1744, which took effect on July 27, 2026, they must take measures to support the development of these competencies. They are no longer required to guarantee a specific level of competency. Shared prompt templates and short training sessions can be one building block.
A good prompt provides the AI with the context it would otherwise lack in the process. In our vision, this doesn’t have to be done manually every time: AI operates within the insurer’s processes and rules, with clearly defined roles and permissions, traceable steps, and humans making the decisions. That’s what we’re working on. It’s not quite there yet, and we’d rather be upfront about that. But those who learn today to describe tasks precisely already have a better understanding of what AI needs to provide reliable assistance in day-to-day insurance operations.
Which tasks would you delegate to AI first, and what should it absolutely never be allowed to decide? We look forward to hearing your thoughts.
Talk to us: tech11.com/en/sales
Sources: Anthropic, “Best Practices for Prompt Engineering” (Claude Blog, 2025); Palantir, “Best Practices for LLM Prompt Engineering” (documentation); State of Berlin, “Guidance on the Use of LLM-Based Chatbots in the State of Berlin” (refers to the DSK guidance “AI and Data Protection” dated May 6, 2024); Regulation (EU) 2016/679 (GDPR), Art. 9; BSI, “Generative AI Models—Opportunities and Risks for Industry and Government Agencies”; Regulation (EU) 2024/1689 (AI Regulation), Art. 4, as amended by Regulation (EU) 2026/1744.